wen hou and cishen zhang - · pdf fileabstract—a novel frame of parallel-beam...
TRANSCRIPT
Abstract—A novel frame of parallel-beam Computerized
Tomography (CT) reconstruction is proposed. The CT
projections are firstly converted to Mojette projections. Then
1D Fourier transform is applied to the modified Mojette
projections, followed by an exact mapping of the gained
Fourier coefficients to the 2D Fourier domain of the scanning
object. Finally the scanning object can be reconstructed from
the partial Fourier coefficients by compressed sensing (CS).
Experimental results show that using this frame, the purpose of
reducing the radiation dosage during CT examinations without
compromising the image quality can be achieved.
Index Terms—Computerized tomography, parallel-beam,
mojette transform, compressed sensing.
I. INTRODUCTION
Computerized Tomography (CT) is a powerful medical
tool which is helpful for the diagnosis of doctors and thus
can benefit patients a lot. In general, the image quality of CT
is proportional to the radiation dosages. But increased
radiation dosages raise the risk of cancer, which becomes an
obstacle for CT development. Hence, it is meaningful and
challenging work to reduce the radiation dosage during CT
examinations without compromising the image quality.
Classical tomography is concerned with the recovery of
the scanning object from a set of projections, which is the
Radon transform (RT) [1] of the object. As Radon transform
is continuous both for the object under scan and the
projections themselves, it has to be sampled to adjust to
practical application, resulting in ill-posedness. Mojette
transform [2] is a both experimentally and computationally
viable discretization of Radon transform. It regularizes the
ill-posedness of inverse Radon transform and allows for an
exact reconstruction in the discrete domain with a finite
number of projections. Despite its advantages, Mojette
transform has two main shortcomings. One is that it is not
compatible with the physical acquisition of CT, and this
problem is addressed in paper [3], in which the authors
design a linear system to calculate the Mojette projections
from a Radon acquisition. The other is that the angle and bin
set is different from practical CT machine, i.e., classical
tomograph acquires are equally distributed over 2π with a
fixed number of bins onto each projection, while Mojette
transform is defined over Farey angles with varying
orientation and number of bins onto each projection. A lot of
research [4] has been done, committed to applying Mojette
Manuscript received October 20, 2012; revised November 24, 2012.
Wen Hou and Cishen Zhang are with the Faculty of Engineering and
Industrial Sciences, Swinburne University of Technology, Vic 3122,
Australia (e-mail: [email protected]; [email protected])
transform to classical tomographic data, which implies the
advantage of Mojette transform on one hand and improves
the potential of Mojette transform in practical application on
the other hand. Paper [4] claims that any set of real, acquired
tomographic data can be rebinned into a compatible Mojette
projection space, without any loss of reconstruction power.
After getting the projections, the next step is the
reconstruction problem. M. Katz [5] presents some
reconstructubility theorems for discrete images and
projections. A lot of work has also been done to recover
object based on Mojette projections, committed to the
improvement of the accuracy. Since we try to reduce the
radiation dosage as well, a newly developed technique
called compressed sensing (CS) is used. CS attempts to
reconstruct signals from significantly fewer samples than
were traditionally thought necessary (e.g., Nyquist sampling
theorem) [6]. CT imagery happens to meet the two key
requirements for successful application of CS: it is
compressible in some transform domain and the scanner
gets encoded samples. Applying CS to CT reconstruction,
the high-accuracy reconstruction with low-dosage radiation
is expected to be obtained [7].
The main contribution of the paper is that a novel
reconstruction frame, i.e. the combination of Mojette
transform and CS, is proposed in the interest of accuracy
and radiation. A detailed description and explanation about
how to modify the Mojette projections for the application of
CS is given. The rest of paper is organized as follows. In
Section II, the Mojette transform and CS is briefly
introduced, and the relation between Mojette transform and
the 2D Fourier coefficients of the scanning object is
formulated, at last the detailed reconstruction algorithm is
described. Section III gives the experimental results to
validate the accuracy, low-dosage radiation and noise
tolerance of the proposed frame. Conclusions are drawn in
Section IV.
II. RECONSTRUCTION FRAME BASED ON MOJETTE
TRANSFORM AND COMPRESSED SENSING
A. Mojette Transform
For simplicity, we assume the scanning object to be
𝑓(𝑥, 𝑦) with the size 𝑁 × 𝑁 (𝑁 is even). CT scanner maps
the 2D object into a set of 1D projection lines, which forms
the sinogram. Each line is the Radon transform of 𝑓 for a
given angle θ and a module ρ [1], defined by:
𝑅𝜃 𝜌 = 𝑓(𝑥, 𝑦)𝛿(𝜌 − 𝑥𝑐𝑜𝑠𝜃 − 𝑦𝑠𝑖𝑛𝜃)+∞
−∞
+∞
−∞𝑑𝑥𝑑𝑦 (1)
Parallel-Beam CT Reconstruction Based on Mojette
Transform and Compressed Sensing
Wen Hou and Cishen Zhang
83DOI: 10.7763/IJCEE.2013.V5.669
International Journal of Computer and Electrical Engineering, Vol. 5, No. 1, February 2013
where θ and ρ are respectively the angular and radial
coordinates of the projection line (𝜃, 𝜌), and 𝛿 is the Dirac's
delta function.
The Dirac-Mojette transform is an exact discretization of
the Radon transform. It is defined over angles =𝑡𝑎𝑛−1(𝑞/𝑝) , where 𝑝 and 𝑞 are relatively prime integers.
As the Farey series of order 𝐾, denoted by 𝐹𝑘 , is the set of
all fractions in lowest terms between 0 and ∞ , whose
denominators don't exceed 𝐾 , e.g.
𝐹4 = [0
1,
1
4,
1
3,
1
2,
2
3,
3
4,
1
1,
4
3,
3
2,
2
1,
3
1,
4
1,
1
0], we can use it to give a
set of discrete angles between [0,𝜋
2] and obtain the angles
over [0, 𝜋] by symmetry. The Dirac-Mojette transform is
defined as:
𝑀𝑝 ,𝑞 𝜌 = 𝑓(𝑥, 𝑦)∆(𝜌 − 𝑝𝑥 − 𝑞𝑦)𝑦𝑥 (2)
The geometry of Dirac-Mojette projector is shown in
Fig.1(a), where the pixel is summed to its corresponding bin
if and only if the X-ray passes through the centre of the
pixel[8].
Fourier Slice Theorem claims that the 1D Fourier
transform of the projections is equal to the 2D Fourier
transform of the image evaluated on the line that the
projection was taken on. In this paper, it is formulated that
the 2D Fourier values of the scanning object can be obtained
exactly through Mojette transform. Let (𝑢, 𝑣) be the
coordinate in the frequency domain, then the 2D discrete
Fourier transform of 𝑓(𝑥, 𝑦) is:
𝐹 𝑢, 𝑣 = 𝑓(𝑥, 𝑦)𝑒−𝑗2𝜋
𝑁(𝑢𝑥 +𝑣𝑦)
𝑦𝑥 (3)
The 1D discrete Fourier transform of 𝑀p,q(𝜌) is:
𝐹𝑀𝑝 ,𝑞 𝑟 = 𝑀𝑝 ,𝑞(𝜌)𝜌 𝑒−𝑗2𝜋
𝐿𝜌𝑟
(4)
where 𝐿 is the projection size, and 𝐿 = 𝑁 − 1 𝑝 +𝑞+1. Combining (2) and (4), we can get:
𝐹𝑀𝑝 ,𝑞 𝑟 = 𝑓(𝑥, 𝑦)𝑒−𝑗2𝜋
𝐿𝑟(𝑝𝑥 +𝑞𝑦 )
𝑦𝑥 (5)
Comparing (3) and (5), it is hard to map 𝐹𝑀𝑝 ,𝑞 𝑟 to
𝐹 𝑢, 𝑣 directly. Considering the periodicity of Fourier
transform, we can merge the items gained by (2) with the
same 𝑚𝑜𝑑(𝑝𝑥 + 𝑞𝑦, 𝑁), then projection size can be reduced
to 𝑁. So
𝑀𝑝 ,𝑞′ 𝜌 = 𝑓(𝑥, 𝑦)|𝜌=𝑚𝑜𝑑 (𝑝𝑥 +𝑞𝑦 ,𝑁)𝑦𝑥 (6)
Its 1D Fourier transform is:
𝐹𝑀𝑝 ,𝑞′ = 𝑓 𝑥, 𝑦 𝑒−𝑗
2𝜋
𝑁𝑟 𝑚𝑜𝑑 𝑝𝑥 +𝑞𝑦 ,𝑁
𝑦𝑥
= 𝑓(𝑥, 𝑦)𝑒−𝑗2𝜋
𝑁[𝑚𝑜𝑑 𝑟𝑝𝑥 +𝑟𝑞𝑦 ,𝑁 ]
𝑦𝑥
(7)
The mapping relationship is revealed over the analysis of
(3) and (7):
∀𝑥, 𝑦, 𝑚𝑜𝑑 𝑢𝑥 + 𝑣𝑦 − 𝑟𝑝𝑥 − 𝑟𝑞𝑦, 𝑁 = 0 (8)
Hence, the 1D Fourier transform of the modified Mojette
projections taken over angles 𝜃 = 𝑡𝑎𝑛−1(𝑞 𝑝 ) can be
mapped to the 2D Fourier plane of the scanning object with
the mapping relationship:
𝑚𝑜𝑑 𝑢 − 𝑟𝑝 = 0 & 𝑚𝑜𝑑 𝑣 − 𝑟𝑞, 𝑁 = 0 (9)
As a result, we can get the partial Fourier coefficients
where 𝑚𝑜𝑑 𝑞𝑢 − 𝑝𝑣, 𝑁 = 0 . The Fourier coefficients
gained through the mapping of the Mojette projections in
Fig. 1. (a) (𝑁 = 6, 𝑡𝑎𝑛𝜃 = 1/2) are marked with star in Fig.
1(b).
(a) Dirac-Mojette Projection of a 6×6 image
(b) The corresponding Fourier values of the Mojette projections
Fig. 1. The geometry of Mojette transform and the mapping relationship in
Fourier domain
B. Compressed Sensing
Compressed sensing is a recently developed theory of
signal recovery from highly incomplete information, and
thus will be employed here to reconstruct the image from
the partial Fourier coefficients. The central idea of CS is that
a sparse or compressible signal 𝑥 ∈ 𝘙𝑁 can be recovered
from a small number of linear measurements 𝑏 = 𝐴𝑥 ∈ 𝘙𝐾 , 𝐾 ≪ 𝑁. The reconstruction can be achieved by solving
the well-known basis pursuit problem [9]-[10]:
min ||𝑥||1 𝑠. 𝑡. 𝐴𝑥 = 𝑏
With the noisy and incomplete samples, an appropriate
relaxation is given by:
min ||𝑥||1 𝑠. 𝑡. | 𝑏 − 𝐴𝑥 |2 ≤ 𝜍
where 𝜍 > 0 is related to the noise.
To gain more effective reconstruction results, a simple
and fast algorithm called RecPF (Reconstruction from
Partial Fourier data) [11] is adopted. It uses an alternating
minimization scheme to solve the following model:
min𝐼 𝑇𝑉 𝐼 + 𝜆||𝜓𝐼||1 + 𝜇||𝐹𝑝 𝐼 − 𝑓𝑝 ||2
where 𝐼 is the image to be reconstructed, 𝑇𝑉(𝐼) is the total
variation regularization term, 𝜓 is a sparsifying basis (e.g.,
wavelet basis), 𝐹𝑝 is a partial Fourier matrix and 𝑓𝑝 denotes
the partial Fourier coefficients. As the main computation of
solving the model only involves shrinkage and fast Fourier
transforms, the reconstruction process is quite fast.
84
International Journal of Computer and Electrical Engineering, Vol. 5, No. 1, February 2013
C. Proposed Frame
The proposed frame contains two stages. At the first
stage, the sinogram is converted to Mojette projections using
the linear system designed in paper [3], then their Fourier
values are mapped to the 2D Fourier domain of the object.
At the second stage, CS is applied to the partial Fourier
coefficients for the recovery of the object. The flowchart is
shown as below.
Fig. 2. Flowchart of the proposed method
Based on the above analysis, it is known that the exact
Fourier coefficients can be obtained through Mojette
transform, which is very beneficial for the reconstruction
accuracy. The application of CS afterwards reduces the
Fourier samples needed for a perfect recovery. Hence, good
reconstruction results can be expected using the proposed
frame when the CT scanning just covers limited number of
view angles.
III. EXPERIMENTAL RESULTS AND DISCUSSION
A. Noise-free reconstruction
In this part, two groups of experiments are designed over
the same Farey angle sets and different Farey angle sets,
respectively. The performance of different methods is
measured through the quantities listed below: the error in the
reconstructed image relative to the original image (Err),
mean square error (MSE) and signal-to-noise ratio (SNR).
Let 𝐼0 denote the original image, 𝐼 the reconstructed image,
then the definitions of the two quantities are:
𝐸𝑟𝑟 = [𝐼 𝑖, 𝑗 − 𝐼0(𝑖, 𝑗)]𝑗𝑖
𝐼0(𝑖, 𝑗)2𝑗𝑖
𝑀𝑆𝐸 = 𝑚𝑒𝑎𝑛[(𝐼 𝑖, 𝑗 − 𝐼0 𝑖, 𝑗 )2]
𝑆𝑁𝑅 = 10 × 𝑙𝑔 (𝐼 𝑖, 𝑗 − 𝑚𝑒𝑎𝑛(𝐼))2
𝑗𝑖
(𝐼 𝑖, 𝑗 − 𝐼0(𝑖, 𝑗))2𝑗𝑖
Firstly, the experiments are conducted over the same
Farey angle sets. We present simulation results of the
256 × 256 phantom image reconstrucion obtained by MCS
and two other methods: FBP and Mift, where FBP uses
filtered backprojection algorithm over the Radon projections
and Mift applies direct inverse Fourier transform to the
Mojette projections. The projections are taken over the
Farey angles. Fig.3 shows the reconstruction results of the
three methods for 𝐹8 projetions, from which we can see that
FBP result has disturbing artifacts due to limited projections,
and Mift can't recover the image from insufficient Fourier
samples. It is obvious that MCS does the almost exact
construction. The Err of the three different methods as a
function of the order of Farey series is plot in Fig.4. We can
see that MCS has already got the good reconstruction result
with 𝐹4 projections, where the Err of Mift and FBP is still
larger than 50%. Fig.4 also shows that FBP can gain better
results than Mift with the increasing number of projections.
(a) original phantom (b) FBP
(c) Mift (d) MCS
Fig. 3. Reconstruction results of the three methods from the projections
taken over F8
Fig. 4. A comparison of Err gained by different methods with the increasing
order of the Farey series
Then the experiments are conducted over different Farey
angle sets. The Mojette filtered backprojection algorithm
(MFBP) [12] is used here for comparison. We use the same
experimental image as MFBP, a 128 × 128 phantom image
consisting of a 17 × 17 square object with unitary value
whereas boundaries are only half valued. Table I lists the
MSE of MCS and MFBP reconstruction results with
different orders of Farey series. We can see that MCS can
gain better results over 𝐹1 angle set with just 4 projection
lines than MFBP over 𝐹32 angle set with 1296 projection
lines. Here are the reconstruction results in Fig.5. It is shown
that the result of MFBP over 𝐹32 angle in (b) has some
disturbing artefacts, while MCS gives clear and exact
85
International Journal of Computer and Electrical Engineering, Vol. 5, No. 1, February 2013
reconstruction result with only 4 projection lines. Fig.5 (d)
shows the line mask of the partial Fourier coefficients which
are the input of CS. If the X-ray tube current is fixed, the
total radiation exposure of the target is proportional to the
number of view angles. Hence, the conclusion is drawn that
less projections are needed for the reconstruction with the
help of CS, which means shorter scanning time and lower
radiation dosage, and thus will benefit the patients more.
(a) original phantom (b) MFBP over 𝐹32 (c) MCS over F1 (d) line mask
Fig. 5. Reconstruction results of MCS and MFBP
TABLE I: MSE OF MCS AND MFBP RECONSTRUCTION RESULTS
Method MCS MFBP
F 𝐹1 𝐹2 𝐹3 𝐹4 𝐹32 𝐹64 𝐹128
#proj 4 8 16 24 1296 5040 20088
MSE 1.69×10-7 6.95×10-8 7.88×10-9 1.23×10-10 0.01322 2×10-5 0
B. Noise tolerance
This experiment is designed to test the noise tolerance of
the proposed frame. We generate our test sets using the
256×256 Shepp-Logan phantom image and FBP is
employed here for comparison. The Gaussian noise is added
to the Mojette projections in MCS and the Radon
projections in FBP, respectively. In both methods, the
projections are taken over F8 angle set, i.e., 88 projections.
As the noise is produced randomly, the quantitative
assessment (Err and SNR) is the mean value of the 20
groups of experimental data, and the results are shown in
Table II. We can see that for the Gaussian noise (0,0.001),
i.e. with mean 0 and standard deviation 0.001, MCS can
gain a much better reconstruction than FBP, the results of
which are shown in Fig.6. But for (0,0.01) noise, the result
of FBP barely changes while the reconstruction quality of
MCS reduces a lot, which means FBP is robust to noise and
MCS is sensitive. The reason is that MCS is based on noise-
sensitive Fourier transform. When (0, 0.01) noise is added to
the projections, the noise of its Fourier coefficients becomes
(0,0.1) for a 256×256 image, verified in Fig.7. Since CS has
certain ability to suppress noise, we get the conclusion that
though sensitive, the proposed frame can deal with small
noise effectively. It will be the future work to deal with the
noisier projections [13].
TABLE II: NOISE RESPONSE OF FBP AND MCS
Noise FBP MCS
Err SNR Err SNR
(0,0.001) 37.1577% 7.3586 2.1929% 31.9393
(0,0.01) 37.1587% 7.3584 23.1656% 11.4631
(a) FBP (b) MCS
Fig. 6. Reconstruction results from the F8 projections with Gaussian noise
(0,0.001)
(a) the noise of projections
(b) the corresponding noise of the Fourier values
Fig. 7. Noise of the projections and the Fourier values
86
International Journal of Computer and Electrical Engineering, Vol. 5, No. 1, February 2013
IV. CONCLUSION
In this paper, a novel frame for parallel-beam CT
reconstruction is presented. Firstly, the sinogram is
converted to the projections gained through Mojette
transform, an exact discretization of Radon transform. On
each view angle, the projections are summed up under some
principles. Then the 1D Fourier coefficients of the merged
projections are mapped to the 2D Fourier domain of the
object. Finally compressed sensing is employed to deal with
the partial Fourier coefficients and can recover the object
very well and suppress the small noise effectively.
Experimental results have demonstrated the advantages of
the proposed method. With the presence of Mojette
transform and compressed sensing, the purpose of reducing
the radiation dosage during CT examinations without
compromising the image quality is achieved.
REFERENCES
[1] P. Toft, “The Radon Transform: Theory and Implementation,” Ph.D.
thesis, Department of Mathematical Modelling, Section for Digital
Signal Processing, Technical University of Denmark, 1996.
[2] J. P. Guedon, D. Barba, and N. Burger, “Psychovisual image coding
via an exact discrete Radon transform,” in Proc. of the SPIE - The
International Society for Optical Engineering, 1995, vol. 2501, pp.
562-572.
[3] B. Recur, P. Desbarats, and J. P. Domenger, “Radon and Mojette
projections’ equivalence for tomographic reconstruction using linear
systems,” in WSCG 2008 Communication Papers, 2008, pp. 191-198.
[4] H. Fayad, J. P. Guedon, I. Svalbe, Y. Bizais, and N. Normand,
“Applying Mojette discrete Radon transforms to classical
tomographic data,” in Proc. of SPIE, 2008, vol. 6913, pp. 1605-7422.
[5] M. Katz, “Questions of Uniqueness and Resolution in Reconstruction
from projections,” Lect. Notes Biomath, vol. 26, Springer-Verlag,
Berlin, Heidelberg, New York, 1979.
[6] M. Lustig, D. L. Donoho, J. M. Santos, and J. M. Pauly, “Compressed
sensing MRI,” IEEE Signal Processing Magazine, pp. 72-82, 2008.
[7] X. Han, J. G. Bian, and D. R. Eaker, “Algorithm-Enabled Low-Dose
Micro-CT Imaging,” IEEE Trans. Medical Imaging, vol. 30, no. 3,
pp. 606-620, 2011.
[8] S. Chandra, N. Normanda, A. Kingston, J. Guedon, and I. Svalbe.
(2010). Fast Mojette transform for discrete tomography. Elsevier
Signal Processing Submitted June (in Review). [Online]. Available:
http://arxiv.org/abs/1006.1965v1
[9] E. Candes, J. Romberg, and T. Tao, “Robust uncertainty principles:
Exact signal reconstruction from highly incomplete frequency
information,” IEEE Trans. Inform. Theory, vol. 52, no. 2, pp. 489-
509, 2006.
[10] D. L. Donoho, Y. Tsaig, I. Drori, and J. L. Starck, “Sparse solution of
underdetermined systems of linear equations by stagewise orthogonal
matching pursuit,” IEEE Trans. Inform. Theory, vol. 58, no. 2, pp.
1094-1121, 2012.
[11] J. F. Yang, Y. Zhang, and W. Yin. A fast TVL1-L2 minimization
algorithm for signal reconstruction from partial Fourier data. IEEE
Journal of Selected Topics in Signal Processing Special Issue on
compressed Sensing. [Online]. Available: http://www.caam.rice.edu/
optimization/L1/RecPF/
[12] M. Servieres, N. Normand, J. P. Guedon, and Y. Bizais, “The Mojette
transform: Discrete angles for tomography,” Electronic Notes in
Discrete Mathematics, vol. 20, pp. 587- 606, 2005.
[13] B. Recur, P. Desbarats, and J. P. Domenger, “Mojette reconstruction
from noisy projections,” in Proc. 2010 2nd International Conference
on Image Processing Theory Tools and Applications (IPTA), 2010, pp.
201-206.
Wen Hou received the B.Eng. degree from Nanjing
University of Aeronautics and Astronautics, China.
She is currently doing Ph.D. in the Faculty of
Engineering and Industrial Sciences, Swinburne
University of Technology, Australia. Her research
interests include medical imaging and
reconstruction.
Cishen Zhang received the B.Eng. degree from
Tsinghua University, China, in 1982 and Ph.D. degree
in Electrical Engineering from Newcastle University,
Australia, in 1990. Between 1971 and 1978, he was an
Electrician with Changxindian (February Seven)
Locomotive Manufactory, Beijing, China. He carried
out research work on control systems at Delft
University of Technology, The Netherlands, from
1983 to 1985. After his Ph.D. study from 1986 to 1989 at Newcastle
University, he was with the Department of Electrical and Electronic
Engineering at the University of Melbourne, Australia as a Lecturer, Senior
Lecturer and Associate Professor and Reader till October 2002. He is
currently with the Faculty of Engineering and Industrial Sciences,
Swinburne University of Technology, Australia. His research interests
include signal processing, medical imaging and control.
87
International Journal of Computer and Electrical Engineering, Vol. 5, No. 1, February 2013